draj
commited on
Commit
•
1f83d09
1
Parent(s):
1766702
All
Browse files- README.md +63 -3
- added_tokens.json +1 -0
- config.json +33 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +1 -0
- spiece.model +3 -0
- spiece.vocab +0 -0
- tokenizer_config.json +1 -0
README.md
CHANGED
@@ -1,3 +1,63 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
This is the IndicBART model. For detailed documentation look here: https://indicnlp.ai4bharat.org/indic-bart/ and https://github.com/AI4Bharat/indic-bart/
|
2 |
+
|
3 |
+
Usage:
|
4 |
+
|
5 |
+
```
|
6 |
+
from transformers import MBartForConditionalGeneration, AutoModelForSeq2SeqLM
|
7 |
+
from transformers import AlbertTokenizer, AutoTokenizer
|
8 |
+
|
9 |
+
tokenizer = AutoTokenizer.from_pretrained("prajdabre/IndicBART", do_lower_case=False, use_fast=False, keep_accents=True)
|
10 |
+
|
11 |
+
# Or use tokenizer = AlbertTokenizer.from_pretrained("prajdabre/IndicBART", do_lower_case=False, use_fast=False, keep_accents=True)
|
12 |
+
|
13 |
+
model = AutoModelForSeq2SeqLM.from_pretrained("prajdabre/IndicBART")
|
14 |
+
|
15 |
+
# Or use model = MBartForConditionalGeneration.from_pretrained("prajdabre/IndicBART")
|
16 |
+
|
17 |
+
# Some initial mapping
|
18 |
+
bos_id = tokenizer._convert_token_to_id_with_added_voc("<s>")
|
19 |
+
eos_id = tokenizer._convert_token_to_id_with_added_voc("</s>")
|
20 |
+
pad_id = tokenizer._convert_token_to_id_with_added_voc("<pad>")
|
21 |
+
# To get lang_id use any of ['<2as>', '<2bn>', '<2en>', '<2gu>', '<2hi>', '<2kn>', '<2ml>', '<2mr>', '<2or>', '<2pa>', '<2ta>', '<2te>']
|
22 |
+
|
23 |
+
# First tokenize the input and outputs. The format below is how IndicBART was trained so the input should be "Sentence </s> <2xx>" where xx is the language code. Similarly, the output should be "<2yy> Sentence </s>".
|
24 |
+
inp = tokenizer("I am a boy </s> <2en>", add_special_tokens=False, return_tensors="pt", padding=True).input_ids # tensor([[ 466, 1981, 80, 25573, 64001, 64004]])
|
25 |
+
|
26 |
+
out = tokenizer("<2hi> मैं एक लड़का हूँ </s>", add_special_tokens=False, return_tensors="pt", padding=True).input_ids # tensor([[64006, 942, 43, 32720, 8384, 64001]])
|
27 |
+
|
28 |
+
model_outputs=model(input_ids=inp, decoder_input_ids=out[:,0:-1], labels=out[:,1:])
|
29 |
+
|
30 |
+
# For loss
|
31 |
+
model_outputs.loss ## This is not label smoothed.
|
32 |
+
|
33 |
+
# For logits
|
34 |
+
model_outputs.logits
|
35 |
+
|
36 |
+
# For generation. Pardon the messiness. Note the decoder_start_token_id.
|
37 |
+
|
38 |
+
model.eval() # Set dropouts to zero
|
39 |
+
|
40 |
+
model_output=model.generate(inp, use_cache=True, num_beams=4, max_length=20, min_length=1, early_stopping=True, pad_token_id=pad_id, bos_token_id=bos_id, eos_token_id=eos_id, decoder_start_token_id=tokenizer._convert_token_to_id_with_added_voc("<2en>"))
|
41 |
+
|
42 |
+
|
43 |
+
# Decode to get output strings
|
44 |
+
|
45 |
+
decoded_output=tokenizer.decode(model_output[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
46 |
+
|
47 |
+
print(decoded_output) # I am a boy
|
48 |
+
|
49 |
+
# What if we mask?
|
50 |
+
|
51 |
+
inp = tokenizer("I am [MASK] </s> <2en>", add_special_tokens=False, return_tensors="pt", padding=True).input_ids
|
52 |
+
|
53 |
+
model_output=model.generate(inp, use_cache=True, num_beams=4, max_length=20, min_length=1, early_stopping=True, pad_token_id=pad_id, bos_token_id=bos_id, eos_token_id=eos_id, decoder_start_token_id=tokenizer._convert_token_to_id_with_added_voc("<2en>"))
|
54 |
+
|
55 |
+
decoded_output=tokenizer.decode(model_output[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
56 |
+
|
57 |
+
print(decoded_output) # I am happy
|
58 |
+
```
|
59 |
+
|
60 |
+
Notes:
|
61 |
+
1. This is compatible with the latest version of transformers but was developed with version 4.3.2 so consider using 4.3.2 if possible.
|
62 |
+
2. While I have only shown how to let logits and loss and how to generate outputs, you can do pretty much everything the MBartForConditionalGeneration class can do as in https://huggingface.co/docs/transformers/model_doc/mbart#transformers.MBartForConditionalGeneration
|
63 |
+
3. Note that the tokenizer I have used is based on sentencepiece and not BPE. Therefore I use the AlbertTokenizer class and not the MBartTokenizer class.
|
added_tokens.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"<s>": 64000, "</s>": 64001, "<2acf>": 64002, "<2eng>": 64003, "<2bis>": 64004, "<2bzj>": 64005, "<2cbk>": 64006, "<2crs>": 64007, "<2djk>": 64008, "<2gul>": 64009, "<2hat>": 64010, "<2hwc>": 64011, "<2icr>": 64012, "<2jam>": 64013, "<2kri>": 64014, "<2ktu>": 64015, "<2mbf>": 64016, "<2mfe>": 64017, "<2mkn>": 64018, "<2pap>": 64019, "<2pcm>": 64020, "<2pis>": 64021, "<2rop>": 64022, "<2sag>": 64023, "<2srm>": 64024, "<2srn>": 64025, "<2tcs>": 64026, "<2tdt>": 64027, "<2tpi>": 64028}
|
config.json
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"activation_dropout": 0.1,
|
3 |
+
"activation_function": "gelu",
|
4 |
+
"architectures": [
|
5 |
+
"MBartForConditionalGeneration"
|
6 |
+
],
|
7 |
+
"attention_dropout": 0.1,
|
8 |
+
"bos_token_id": 64000,
|
9 |
+
"d_model": 1024,
|
10 |
+
"classifier_dropout": 0.0,
|
11 |
+
"decoder_attention_heads": 16,
|
12 |
+
"decoder_ffn_dim": 4096,
|
13 |
+
"decoder_layerdrop": 0.0,
|
14 |
+
"decoder_layers": 6,
|
15 |
+
"dropout": 0.1,
|
16 |
+
"encoder_attention_heads": 16,
|
17 |
+
"encoder_ffn_dim": 4096,
|
18 |
+
"encoder_layerdrop": 0.0,
|
19 |
+
"encoder_layers": 6,
|
20 |
+
"eos_token_id": 64001,
|
21 |
+
"gradient_checkpointing": false,
|
22 |
+
"init_std": 0.02,
|
23 |
+
"is_encoder_decoder": true,
|
24 |
+
"max_position_embeddings": 1024,
|
25 |
+
"model_type": "mbart",
|
26 |
+
"num_hidden_layers": 6,
|
27 |
+
"pad_token_id": 0,
|
28 |
+
"scale_embedding": false,
|
29 |
+
"transformers_version": "4.3.2",
|
30 |
+
"use_cache": true,
|
31 |
+
"vocab_size": 64015,
|
32 |
+
"tokenizer_class": "AlbertTokenizer"
|
33 |
+
}
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8a8329f70aacb70d9eb786587f35b1ef6f55e487fb1039e653a2e869c15e658b
|
3 |
+
size 976483953
|
special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"bos_token": "[CLS]", "eos_token": "[SEP]", "unk_token": "<unk>", "sep_token": "[SEP]", "pad_token": "<pad>", "cls_token": "[CLS]", "mask_token": {"content": "[MASK]", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true}, "additional_special_tokens": ["<s>", "</s>", "<2acf>", "<2eng>", "<2bis>", "<2bzj>", "<2cbk>", "<2crs>", "<2djk>", "<2gul>", "<2hat>", "<2hwc>", "<2icr>", "<2jam>", "<2kri>", "<2ktu>", "<2mbf>", "<2mfe>", "<2mkn>", "<2pap>", "<2pcm>", "<2pis>", "<2rop>", "<2sag>", "<2srm>", "<2srn>", "<2tcs>", "<2tdt>", "<2tpi>"]}
|
spiece.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5a2905611b144e8ec04764787aea304cda4981550b3ec545f5a479264f1471c5
|
3 |
+
size 1348913
|
spiece.vocab
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"do_lower_case": false, "remove_space": true, "keep_accents": true, "bos_token": "[CLS]", "eos_token": "[SEP]", "unk_token": "<unk>", "sep_token": "[SEP]", "pad_token": "<pad>", "cls_token": "[CLS]", "mask_token": {"content": "[MASK]", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "use_fast": false, "strip_accents": false, "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "/share03/draj/data/parallel_corpora/creole/jw300/splits/vocabs/albert-all64k"}
|